EXAONE 3.5 8B

Positioning

EXAONE 3.5 8B is a dense 7.8B-parameter language model released by LG AI Research under the EXAONE License. With a 32,768-token context window, it is designed for consumer-tier Korean and broader CJK workloads. As a smaller, dense model, it offers straightforward deployment on consumer hardware without the complexity of mixture-of-experts architectures.

Strengths

• Optimized for Korean/CJK workloads: Built by LG AI Research with a focus on Korean language tasks, making it a strong candidate for applications requiring high-quality Korean text generation and understanding.
• Consumer-friendly size: At 7.8B parameters, the model fits comfortably on a single consumer GPU, with quantized versions as small as ~2.5 GB (Q2_K) on disk.
• Permissive commercial license: The EXAONE License allows commercial use, making it suitable for businesses deploying Korean-language AI solutions locally.
• Long context window: 32,768 tokens of context enable processing of substantial documents or multi-turn conversations without truncation.

Limitations

• Narrow language focus: While strong for Korean, its performance in other languages may be less competitive; operators targeting multilingual workloads should evaluate carefully.
• No community benchmarks yet: We do not have independent, community-reported benchmark results for this model. Published vendor metrics should be treated as best-case.
• Dense architecture: Unlike MoE models that offer high parameter counts with lower inference cost, this dense 7.8B model requires full compute per token, limiting throughput on lower-end hardware.
• Limited ecosystem: As a relatively new model from a non-frontier lab, community tooling, fine-tuning recipes, and third-party quantizations may be less mature than for more popular open-weight models.

What it takes to run this locally

At FP16, the model requires 16 GB of disk space and roughly 16 GB of VRAM, plus additional memory for KV cache and framework overhead (typically 30–50% more at full context). Quantized versions reduce requirements significantly: Q4_K_M (4.4 GB) fits on most 8 GB GPUs, while Q2_K (~2.5 GB) can run on 4–6 GB cards. This places the model firmly in the consumer deployment class, suitable for single GPUs with 8–24 GB VRAM.

Should you run this locally?

Yes if you need a locally deployable, commercially licensed model optimized for Korean-language tasks and have a consumer GPU with at least 8 GB VRAM for quantized inference.

No if your primary workload is English or multilingual, or if you require extensive community support and pre-built integrations. In those cases, more established open-weight models may be a better fit.

Catalog cross-links

• EXAONE 3.0 7.8B
• EXAONE 3.5 32B